Transmission Control Protocol (TCP) carries the bulk of all traffic
across most common types of data networks, including those in today's
high-speed, low-latency datacenters. A key problem faced by commodity
datacenters is incast congestion. Client queries are distributed by
front-end nodes over TCP connections to multiple back-end servers, and
the replies coming back tend to be highly correlated in time. This
correlated reply traffic causes microsecond-timescale congestion (which
we call microburst congestion
in the multi-gigabit Ethernet switch buffers along the return path.
TCP reacts far too slowly to packet losses in the context of high-speed
and low RTT networks. Consequently, any packet loss caused by incast
(microburst) congestion will stall the higher layers' data gathering
pipeline (potentially for many tens of milliseconds) and unnecessarily
increase client response times as a result.
We propose to evaluate emerging reactive (e.g. faster control loops)
and proactive (e.g. burst de-correlation) techniques for managing
incast TCP congestion events.
- To identify and statistically characterise microburst
congestion events common in data centres
- To create a real and/or simulation testbed capable of
emulating different types of microburst congestion events
- To empirically evaluate at least one (and possibly
more) TCP modification(s) for reducing (and hopefully eliminating) the
occurrence of TCP-triggered microburst congestion events in the
emulated testbed environment.
In September 2013 we released a FreeBSD9-based VirtualBox VM appliance
containing NS-3 and a version of NSC (Network simulation cradle)
patched to work with the actual TCP stacks from FreeBSD9 and
FreeBSD-HEAD. This enables researchers to perform simulations of incast
topologies using the same TCP algorithms as implemented in real FreeBSD
This project has been made possible in
part by a gift from The
Cisco University Research Program Fund, a corporate advised
fund of Silicon Valley Community Foundation, for a project titled
"Exploring possible mitigation for incast TCP congestion in data centres".